[2603.09906] Thinking to Recall: How Reasoning Unlocks Parametric Knowledge in LLMs

Reasoning helps LLMs recall simple factual knowledge via a computational buffer effect and factual priming, but hallucinated intermediate facts increase final errors.